Notes - When AI Meets SaaS: Disruption, Adaptation, and the Moat of Taste
Summary
- AI threatens enterprise SaaS more than consumer apps, as agentic interfaces and in-house AI development erode seat-based pricing and standardized workflows while incumbents survive on backend depth, governance, and trust.
- The “taste” factor—human intuition in design, vision, and product adoption—remains AI’s weakest area, protecting consumer tech and creative domains from rapid automation.
- While AI can cheaply generate code, maintaining, refining, and integrating software still carries massive hidden costs, ensuring that engineering expertise and architecture remain the real value drivers.
- Enterprise software’s dominance has long depended on sales-led adoption and executive persuasion rather than pure product merit—an inefficiency AI procurement agents could eventually disrupt.
- Across subsectors—from stablecoins and adtech to design, data, and chip tools—AI’s impact will be uneven: commoditizing low-complexity layers while reinforcing moats built on integration, human “taste,” and specialized trust.
- We review AI disruption threats to a number of companies, including Circle, Palantir, Cloudflare, Figma, Snowflake, Oracle, Cadence, and Synopsis.
There is much speculation as to how AI-native applications are going to impact the incumbent software world. We've discussed this topic already in two previous notes.
Notes: Navigating the Shifting Sands: Is Enterprise Software Facing a Secular Sentiment Change?
Notes: Incumbents vs. AI-Natives: Who Wins the Next Software Cycle?
Some voices in the market have argued that AI could disrupt SaaS, allowing users to build custom applications more closely aligned with their unique needs. This view often reduces software to little more than a CRUD database layered with business logic.
However, a more prevailing perspective — and the one we see as more realistic — is that SaaS remains indispensable because of its integrated data, governance, and application layers. These elements provide value that individual AI-generated applications cannot easily replicate.
The Debate: Is AI Really a Threat to SaaS?
Recently, the debate over whether AI will disrupt the software application industry has intensified. This is due in part to the remarkable progress made by AI coding agents and the increasing "real-world applicability" of AI models — moving beyond just impressive test scores on benchmarks like MMLU.
Proponents of AI disruption present compelling arguments. In contrast, defenders of the status quo — often led by established players like Salesforce — seem to offer weaker counterpoints.
If the AI disruptors are correct, we would expect to see consumers and product managers, empowered by no-code or low-code AI tools, reshaping the consumer tech landscape first. However, the most dramatic impact has actually been seen in enterprise software, with a significant "bloodbath" in the sector throughout the first and second halves of 2025.
Why Enterprise SaaS is Most Vulnerable
Several factors contribute to this vulnerability:
- Seat-based Pricing is Exposed: SaaS companies that charge per user are particularly at risk. AI-driven efficiency gains reduce headcount, and AI agents can even replace human roles entirely — undermining the traditional pricing model.
- Enterprises Can Build Their Own: Unlike individual consumers, enterprises have the budget and technical resources to leverage AI in creating custom software. This puts pressure on traditional enterprise SaaS applications, which are often expensive and overly standardized, making them vulnerable to being replaced by AI-driven, in-house alternatives.
- Budget Shift to AI Projects: AI initiatives are absorbing new budget allocations and even cannibalizing existing enterprise software budgets.
- More Opportunity to Improve Upon Enterprise SaaS: There are tens of domains in running an enterprise, and within each domain there are hundreds of processes and workflows. This offered SaaS-natives great opportunities to improve upon legacy on-prem software throughout the 2010s. Now, the plethora of different enterprise work offers AI-natives great opportunities to improve upon legacy SaaS.
All these trends are converging, leading us to believe that only a handful of SaaS companies will survive what could be described as an "Enterprise Software Extinction Event." However, those survivors — such as Palantir (PLTR), Snowflake (SNOW), and Oracle (ORCL) — are likely to thrive, as they are actively embracing AI and are less vulnerable to disruption due to their sophisticated, human-curated logic.
Caveats and the Path Forward
Still, it is important to note a few caveats to Sam Altman’s narrative that "enterprise software is finished." Our own experience in building AI-native applications suggests there is more nuance to the story.
Notably, Anthropic recently released an experimental product, Imagine with Claude, and OpenAI published a paper showing GPT achieving expert-level performance in top industries. These developments hint at a future where the line between AI and traditional SaaS may blur, but not disappear entirely.
Imagine with Claude is basically a virtual desktop environment where users can start via describing what they need, and the software will be generated and modified in real-time as opposed to using pre-made software.
https://openai.com/index/gdpval/
GPT and Claude are getting closer to humans' actual performance in real-life industry-specific tasks. Note that this is based on one-shot comparison between humans and AI. For real world interactive multi-round tasks, we are still possibly many quarters away from hitting the tipping point.
Taste - UI/UX/Full stack
The Role of “Taste” in UI/UX and Full Stack Development
So, where should we begin when evaluating the future of software and AI? Surprisingly, the biggest limitation users face when interacting with LLMs often isn’t the model’s output quality — it’s the quality of the input provided by users. In many cases, dissatisfaction with AI-generated applications arises not from the inherent limitations of the technology, but from the inability of users to ask the right, high-quality questions or provide clear direction.
For example, if you ask an AI to design “a great smartphone, like the iPhone,” you’re unlikely to receive a result that matches the iPhone’s renowned polish and sophistication. Instead, you’ll get a basic prototype — decent, perhaps, but nowhere near the level of excellence set by Apple’s design team. This gap underscores a critical point: in software and product development, taste matters — a lot.
Taste — the elusive ability to create products that are not just functional but delightful — matters most in consumer technology. This helps explain why consumer apps face less disruption from AI: users expect seamless experiences, design intuition, and emotional resonance that AI still struggles to replicate.
Enterprise SaaS and the Role of AI
In enterprise software, taste is far less critical. Applications are built for functionality, not delight, which makes them easier to disintermediate with AI. What threatens SaaS most here are not AI coding agents per se, but AI agentic interfaces that can sit on top of data and workflows, bypassing the SaaS UI entirely. These agentic layers — often built with the help of AI coding tools — let users interact with enterprise systems through natural language rather than pre-built dashboards and menus.
However, enterprise SaaS incumbents retain powerful moats. Their strength lies in backend integrations, security, governance, and above all their status as trusted custodians of the System of Record. Even if the UI layer is replaced by agents, enterprises will still rely on the underlying platforms to safeguard data and ensure compliance.
Where AI Excels and Where It Falls Short
AI coding agents excel at generating interfaces and workflow scaffolding, but complex backend engineering and cross-system integrations remain difficult. Meanwhile, agentic interfaces shine in usability — replacing rigid UIs with natural, conversational ones. This creates real pressure on SaaS vendors whose differentiation is mostly in their front-end experience, but less so on those whose value rests on backend depth and data stewardship.
At the frontier of most domains, even investing, it seems human intuition will still drive innovation forward. To generate text or code, AI models apply their knowledge accrued via their training on public information to make a very good guess at what the next token they generate should be. As a result, if AI were to develop a thesis for a stock, it would generate a thesis that is common among the investing community of that company. However, we all know that the best returns derive from having an uncommon thesis, not a common one. Likewise, while AI models are great at putting together a list of known malware and instructions for a hacker to launch a successful attack, they cannot currently develop entirely new malware or attack types. The lack of "thinking out the box" within AI models is what gives humans an edge in the frontier of most domains, especially creative ones that require taste, like product design.
